Don't build punica kernels by default (vllm-project#2605)

.github/workflows/scripts/build.sh

@@ -13,6 +13,8 @@ $python_executable -m pip install -r requirements.txt
 
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1
+# Make sure punica is built for the release (for LoRA)
+export VLLM_INSTALL_PUNICA_KERNELS=1
 
 # Build
 $python_executable setup.py bdist_wheel --dist-dir=dist

Dockerfile

@@ -45,6 +45,8 @@ ENV MAX_JOBS=${max_jobs}
 # number of threads used by nvcc
 ARG nvcc_threads=8
 ENV NVCC_THREADS=$nvcc_threads
+# make sure punica kernels are built (for LoRA)
+ENV VLLM_INSTALL_PUNICA_KERNELS=1
 
 RUN python3 setup.py build_ext --inplace
 #################### EXTENSION Build IMAGE ####################

setup.py

@@ -265,7 +265,7 @@ def get_torch_arch_list() -> Set[str]:
         with contextlib.suppress(ValueError):
             torch_cpp_ext.COMMON_NVCC_FLAGS.remove(flag)
 
-    install_punica = bool(int(os.getenv("VLLM_INSTALL_PUNICA_KERNELS", "1")))
+    install_punica = bool(int(os.getenv("VLLM_INSTALL_PUNICA_KERNELS", "0")))
     device_count = torch.cuda.device_count()
     for i in range(device_count):
         major, minor = torch.cuda.get_device_capability(i)

vllm/lora/punica.py

@@ -157,10 +157,13 @@ def _raise_exc(
         **kwargs  # pylint: disable=unused-argument
     ):
         if torch.cuda.get_device_capability() < (8, 0):
-            raise ImportError(
-                "LoRA kernels require compute capability>=8.0") from import_exc
+            raise ImportError("punica LoRA kernels require compute "
+                              "capability>=8.0") from import_exc
         else:
-            raise import_exc
+            raise ImportError(
+                "punica LoRA kernels could not be imported. If you built vLLM "
+                "from source, make sure VLLM_INSTALL_PUNICA_KERNELS=1 env var "
+                "was set.") from import_exc
 
     bgmv = _raise_exc
     add_lora = _raise_exc